Imagine waking up to 722 new maths papers. Not 722 tweets about papers — 722 manuscripts, organised into 372 “result families,” posted to a public repository in a single upload on October 6, 2026.
That's what OpenAI did. And almost immediately, the internet did what the internet does: it rounded “722 manuscripts” up to “700 problems solved.”
Those are not the same sentence. A manuscript is a claim with working shown. A solved problem is a claim other mathematicians have checked, stress-tested, and accepted. Between those two things lies peer review, formal checking, and time — and for most of the 722, that work hasn't happened yet.
What “formalized” actually means here
Some of the results come with a Lean formalization — a version of the proof written so a computer can check each logical step. That's genuinely new at this scale, and it's a big part of why this drop matters.
But even the numbers disagree with each other. OpenAI's catalogue says roughly 42% are formalized. An independent count lands closer to 22% — about 162 of the 722. And within two days, three manuscripts were withdrawn (a sign error broke one, and two depended on it) while fourteen more were repaired. That doesn't make the drop fake. It makes it a first draft at industrial scale.
The headline claims inside the batch
Even with the caveats, the list of claimed results is startling. Treat every item below as “claimed in the drop,” not “settled”:
- Quasi-Riemann hypothesis — no Dirichlet L-function zeros with Re(s) > 7/8, with the Landau–Siegel zero eliminated. Lean-formalized per the repo, with a human-polished write-up. This is not the Millennium Riemann Hypothesis.
- Rational Hodge conjecture for CM abelian varieties — a special case, with no Lean formalization.
- BSD leading-term formula — for special Selmer cases only.
- Plus claimed results on the Unique Games Conjecture, Mahler conjectures, a Kaplansky zero-divisor disproof, L = RL = BPL, a Hadwiger disproof, the 4D Kakeya maximal conjecture, Hilbert's tenth problem over ℚ, and a 9/4 matrix-multiplication exponent bound.
Any one of those, confirmed, would be a career headline. All of them arriving in one upload is exactly why verification — not announcement — is now the bottleneck in maths.
So what is the actual story?
Not “AI solved 700 problems.” It's this: AI can now produce plausible, sometimes machine-checked mathematics faster than humans can verify it. Scale arrived before checking did.
That's still a historic shift. It just isn't 722 settled solutions — and Slopdrop won't call them that until the checking catches up.